Crystal Identification in DOI-PET Detector Using SOM Neural Network and Mean-shift Algorithm

The positron emission tomography (PET) is a nuclear medicine device for molecular, metabolic and functional imaging, which is extensively used in nuclear medicine for clinical examination and preclinical research. The key component of a PET device is the gammaray detectors, which commonly consist of scintillator arrays coupled to photon sensor arrays. This type of detector needs to segment its flood source image to generate a crystal position lookup table (LUT). The accuracy of the LUT is critical to the system performance. For a whole PET system, the number of detector blocks may be hundreds, thus it will be time consuming if the process is done by manual segmentation. An automatic algorithm for the crystal recognition and segmentation of flood maps generated by an animal PET system with depth of interaction (DOI) capability based on 48 duallayeroffset detector blocks was proposed in this paper. The top and bottom layers were directly distinguished using the intensity difference and offset grid pattern. The identification of the response peaks of the top layer was based on the singular value decomposition (SVD) and meanshift algorithm. SVD was employed to create a principal component image of the top layer. Then, projection profiles along the x and y directions are obtained. A local maximum identification method was utilized to locate the peaks from these projections. At last, the meanshift algorithm was used to improve the accuracy of the peaks. Identification of the response peaks of the bottom layer was based on selforganizing map (SOM) neural networks and meanshift algorithm. Initial peaks of the bottom layer were generated based on the shift of the top peaks. Then they were adjusted using the SOM algorithm simultaneously. At last, they were modified individually using the meanshift algorithm. After locating all response peaks, the flood map was segmented using an Euclidean distance based algorithm. The proposed algorithm was run on a laptop with the Intel i56300@2.30GHz CPU for the whole PET system. The results show that it achieves a crystal peak identification accuracy of 99.56% for the top layer and 99.11% for the bottom layer, the average accuracy of the whole system is 99.34%. The average processing time for a block based on the laptop is 101 s. Compared with the algorithm with only meanshift algorithm, the SOM algorithm improves the identification quality for the bottom layer. In conclusion, a robust, fast, high accuracy crystal identification method for duallayeroffset DOIPET detectors are developed. The proposed method can also be utilized for single layer PET detector blocks.